Hyperspectral image classification with multivariate empirical mode decomposition-based features

Previous studies have demonstrate that the empirical mode decomposition (EMD) can provide significant improvements in hyperspectral classification due to its ability to extract the nature scale components (i.e. intrinsic mode functions (IMFs)) of the hyperspectral image (HSI) adaptively. However, the IMFs gained from various hyperspectral bands may be different in number and frequency, heavily compromising the analysis of HSI obtained in a channel-by-channel basis. To cope with this problem, we utilize the multivariate EMD (MEMD), for the first time, in HSI classification. Core steps of the proposed method are threefold: 1) appropriate bands from the original HSI are selected by a mutual-information-based way to mitigate the 'curse of dimensionality'; 2) each of the selected bands is vectorized into a row vector. All the row vectors obtained from the selected bands are then combined to form different part of a multivariate signal, which can be decomposed by the MEMD; 3) the generated features (i.e. sum of the IMFs) are finally classified by the widely used support vector machine (SVM). Experiments on the benchmark Indian Pines data demonstrate the feasibility of the proposed method in enhancing the classification performance, making it highly promising for further study.